Splitting a model across devices buys memory and spends interconnect

Tensor parallelism reduces some per-device memory costs while adding communication between devices. Calculate both, check which KV state is replicated, and compare replica layouts using the same total device budget and latency target.

15 MIN · PREMIUM